There was an issue I mentioned back in episode 4. Gemini 3.0 Pro kept not following the project guidelines I'd written. It would ignore the style guide, skip the component structure I'd defined, and repeatedly break rules I'd spelled out in GEMINI.md. The workflows I'd set up as global skills didn't work as expected either. At first I assumed it would improve if I rewrote prompts more precisely or refined the guidelines. But after trying for over a month, nothing got better. That's when I decided to switch AI tools.
Why I Chose Claude Code
There wasn't much to deliberate about. Among AI coding tools for development, Claude Code was the one I judged to be most effective, and since my company already used it, I was reasonably familiar. I didn't have to absorb the cost of learning an entirely unfamiliar tool. The real questions were how to structure the new guidelines system and how to migrate the content from GEMINI.md — not how to learn the tool itself.
The Biggest Change: It Actually Follows Instructions
The biggest difference I felt after switching is straightforward. It follows instructions. It doesn't ignore the rules written in CLAUDE.md, it respects the style guide, and it doesn't stray from the established component usage patterns. When I was using Gemini, I often had to make the same corrections during code review over and over. That frequency dropped noticeably. It's not perfect and occasionally misses something, but there's no comparison to before.
Subagents: Main Manages, Work Gets Delegated
Switching to Claude Code opened up the concept of subagents. When I was using Gemini, all work happened within a single conversation thread, and as tasks got longer, earlier context would sometimes get lost. When that happened, I had to re-explain the background, which was a nuisance.
Multiple independent tasks can also run in parallel. For example, when I need to create several psychology test data files at once, I can spin up multiple subagents and split the work between them simultaneously. Before, I processed them one at a time in sequence. Now that time is significantly reduced.
Connected Slack Notifications via Hooks
Moving to Claude Code also meant I could use the hooks feature. Hooks are essentially scripts that Claude Code runs automatically at specific points. When a task finishes, or when it's waiting for input, a designated action runs automatically. I connected this to Slack notifications. When a task completes or my input is needed, I get a Slack message.
My actual working efficiency improved. Before, I had to go check whether a task was done after assigning it. Since I couldn't gauge how long things would take, I kept checking in between, which disrupted my focus. Now I can concentrate on something else until the Slack notification arrives. It seems simple, but the practical difference is noticeable.
While I Was at It, I Overhauled the Documentation Too
Switching AI tools meant I had to migrate the content from GEMINI.md into a new structure. Since I had to touch it anyway, I decided to clean up the documentation overall. Looking back, there was duplicate content scattered across files, and some content had gotten stale and no longer matched the actual code. Some documents had too much mixed together, making it hard to find specific information.
- Split by role: Instead of one long file, I separated files by workflow and topic. The file to reference for adding a new psychology test, adding a mini-game, or updating documentation is now clearly distinct for each situation.
- Removed duplication: Content that was written identically across multiple documents was consolidated in one place, with other documents referencing it. When something needs updating later, I only have to change it in one spot.
- Added situation-specific guides: I put together a table so it's clear at a glance which document to consult for each type of task. Agents can navigate directly without having to explore from scratch every time.
It took more time than I expected. Organizing documentation might seem less important than writing code, but with an AI agent workflow, documentation quality directly correlates to output quality. Documentation is what tells the agent what to do and how. I feel the overall quality improved considerably.
One Regret: Cost
I've talked about a lot of the positives, but that doesn't mean there are no downsides. The Pro plan quota runs out too fast. A focused work session and you hit the limit quickly. So I'm now on the Max plan, which costs considerably more. For a side project, it's a notable expense.
It's clear that the workflow runs more stably since switching tools. Instruction compliance, improved context management through subagents, and the breathing room created by Slack notifications — these three things made a real difference. The cost issue remains, but the next goal is for Q-Fit to grow enough to solve it.